Kris M. Kitani

Qichen Fu, Xingyu Liu, Ran Xu et al. · cmu, salesforce

Accurately estimating 3D hand pose is crucial for understanding how humans interact with the world. Despite remarkable progress, existing methods often struggle to generate plausible hand poses when the hand is heavily occluded or blurred. In videos, the movements of the hand allow us to observe various parts of the hand that may be occluded or blurred in a single frame. To adaptively leverage the visual clue before and after the occlusion or blurring for robust hand pose estimation, we propose the Deformer: a framework that implicitly reasons about the relationship between hand parts within the same image (spatial dimension) and different timesteps (temporal dimension). We show that a naive application of the transformer self-attention mechanism is not sufficient because motion blur or occlusions in certain frames can lead to heavily distorted hand features and generate imprecise keys and queries. To address this challenge, we incorporate a Dynamic Fusion Module into Deformer, which predicts the deformation of the hand and warps the hand mesh predictions from nearby frames to explicitly support the current frame estimation. Furthermore, we have observed that errors are unevenly distributed across different hand parts, with vertices around fingertips having disproportionately higher errors than those around the palm. We mitigate this issue by introducing a new loss function called maxMSE that automatically adjusts the weight of every vertex to focus the model on critical hand parts. Extensive experiments show that our method significantly outperforms state-of-the-art methods by 10%, and is more robust to occlusions (over 14%).

Takehiko Ohkawa, Yu-Jhe Li, Qichen Fu et al. · cmu

We aim to improve the performance of regressing hand keypoints and segmenting pixel-level hand masks under new imaging conditions (e.g., outdoors) when we only have labeled images taken under very different conditions (e.g., indoors). In the real world, it is important that the model trained for both tasks works under various imaging conditions. However, their variation covered by existing labeled hand datasets is limited. Thus, it is necessary to adapt the model trained on the labeled images (source) to unlabeled images (target) with unseen imaging conditions. While self-training domain adaptation methods (i.e., learning from the unlabeled target images in a self-supervised manner) have been developed for both tasks, their training may degrade performance when the predictions on the target images are noisy. To avoid this, it is crucial to assign a low importance (confidence) weight to the noisy predictions during self-training. In this paper, we propose to utilize the divergence of two predictions to estimate the confidence of the target image for both tasks. These predictions are given from two separate networks, and their divergence helps identify the noisy predictions. To integrate our proposed confidence estimation into self-training, we propose a teacher-student framework where the two networks (teachers) provide supervision to a network (student) for self-training, and the teachers are learned from the student by knowledge distillation. Our experiments show its superiority over state-of-the-art methods in adaptation settings with different lighting, grasping objects, backgrounds, and camera viewpoints. Our method improves by 4% the multi-task score on HO3D compared to the latest adversarial adaptation method. We also validate our method on Ego4D, egocentric videos with rapid changes in imaging conditions outdoors.

Xingyu Liu, Deepak Pathak, Kris M. Kitani · cmu

The ability to learn from human demonstration endows robots with the ability to automate various tasks. However, directly learning from human demonstration is challenging since the structure of the human hand can be very different from the desired robot gripper. In this work, we show that manipulation skills can be transferred from a human to a robot through the use of micro-evolutionary reinforcement learning, where a five-finger human dexterous hand robot gradually evolves into a commercial robot, while repeated interacting in a physics simulator to continuously update the policy that is first learned from human demonstration. To deal with the high dimensions of robot parameters, we propose an algorithm for multi-dimensional evolution path searching that allows joint optimization of both the robot evolution path and the policy. Through experiments on human object manipulation datasets, we show that our framework can efficiently transfer the expert human agent policy trained from human demonstrations in diverse modalities to target commercial robots.

Zhengyi Luo, Ye Yuan, Kris M. Kitani · cmu

Automatically designing virtual humans and humanoids holds great potential in aiding the character creation process in games, movies, and robots. In some cases, a character creator may wish to design a humanoid body customized for certain motions such as karate kicks and parkour jumps. In this work, we propose a humanoid design framework to automatically generate physically valid humanoid bodies conditioned on sequence(s) of pre-specified human motions. First, we learn a generalized humanoid controller trained on a large-scale human motion dataset that features diverse human motion and body shapes. Second, we use a design-and-control framework to optimize a humanoid's physical attributes to find body designs that can better imitate the pre-specified human motion sequence(s). Leveraging the pre-trained humanoid controller and physics simulation as guidance, our method is able to discover new humanoid designs that are customized to perform pre-specified human motions.

Xiaofang Wang, Kris M. Kitani · cmu

Considerable research effort has been devoted to LiDAR-based 3D object detection and empirical performance has been significantly improved. While progress has been encouraging, we observe an overlooked issue: it is not yet common practice to compare different 3D detectors under the same cost, e.g., inference latency. This makes it difficult to quantify the true performance gain brought by recently proposed architecture designs. The goal of this work is to conduct a cost-aware evaluation of LiDAR-based 3D object detectors. Specifically, we focus on SECOND, a simple grid-based one-stage detector, and analyze its performance under different costs by scaling its original architecture. Then we compare the family of scaled SECOND with recent 3D detection methods, such as Voxel R-CNN and PV-RCNN++. The results are surprising. We find that, if allowed to use the same latency, SECOND can match the performance of PV-RCNN++, the current state-of-the-art method on the Waymo Open Dataset. Scaled SECOND also easily outperforms many recent 3D detection methods published during the past year. We recommend future research control the inference cost in their empirical comparison and include the family of scaled SECOND as a strong baseline when presenting novel 3D detection methods.

Xingyu Liu, Deepak Pathak, Kris M. Kitani

A popular paradigm in robotic learning is to train a policy from scratch for every new robot. This is not only inefficient but also often impractical for complex robots. In this work, we consider the problem of transferring a policy across two different robots with significantly different parameters such as kinematics and morphology. Existing approaches that train a new policy by matching the action or state transition distribution, including imitation learning methods, fail due to optimal action and/or state distribution being mismatched in different robots. In this paper, we propose a novel method named $REvolveR$ of using continuous evolutionary models for robotic policy transfer implemented in a physics simulator. We interpolate between the source robot and the target robot by finding a continuous evolutionary change of robot parameters. An expert policy on the source robot is transferred through training on a sequence of intermediate robots that gradually evolve into the target robot. Experiments on a physics simulator show that the proposed continuous evolutionary model can effectively transfer the policy across robots and achieve superior sample efficiency on new robots. The proposed method is especially advantageous in sparse reward settings where exploration can be significantly reduced. Code is released at https://github.com/xingyul/revolver.

Shun Iwase, Xingyu Liu, Rawal Khirodkar et al.

We present RePOSE, a fast iterative refinement method for 6D object pose estimation. Prior methods perform refinement by feeding zoomed-in input and rendered RGB images into a CNN and directly regressing an update of a refined pose. Their runtime is slow due to the computational cost of CNN, which is especially prominent in multiple-object pose refinement. To overcome this problem, RePOSE leverages image rendering for fast feature extraction using a 3D model with a learnable texture. We call this deep texture rendering, which uses a shallow multi-layer perceptron to directly regress a view-invariant image representation of an object. Furthermore, we utilize differentiable Levenberg-Marquardt (LM) optimization to refine a pose fast and accurately by minimizing the feature-metric error between the input and rendered image representations without the need of zooming in. These image representations are trained such that differentiable LM optimization converges within few iterations. Consequently, RePOSE runs at 92 FPS and achieves state-of-the-art accuracy of 51.6% on the Occlusion LineMOD dataset - a 4.1% absolute improvement over the prior art, and comparable result on the YCB-Video dataset with a much faster runtime. The code is available at https://github.com/sh8/repose.

Shengcao Cao, Xiaofang Wang, Kris M. Kitani

We propose a method to incrementally learn an embedding space over the domain of network architectures, to enable the careful selection of architectures for evaluation during compressed architecture search. Given a teacher network, we search for a compressed network architecture by using Bayesian Optimization (BO) with a kernel function defined over our proposed embedding space to select architectures for evaluation. We demonstrate that our search algorithm can significantly outperform various baseline methods, such as random search and reinforcement learning (Ashok et al., 2018). The compressed architectures found by our method are also better than the state-of-the-art manually-designed compact architecture ShuffleNet (Zhang et al., 2018). We also demonstrate that the learned embedding space can be transferred to new settings for architecture search, such as a larger teacher network or a teacher network in a different architecture family, without any training. Code is publicly available here: https://github.com/Friedrich1006/ESNAC .

Rohan Choudhury, Guanglei Zhu, Sihan Liu et al.

Transformers are slow to train on videos due to extremely large numbers of input tokens, even though many video tokens are repeated over time. Existing methods to remove such uninformative tokens either have significant overhead, negating any speedup, or require tuning for different datasets and examples. We present Run-Length Tokenization (RLT), a simple approach to speed up video transformers inspired by run-length encoding for data compression. RLT efficiently finds and removes runs of patches that are repeated over time prior to model inference, then replaces them with a single patch and a positional encoding to represent the resulting token's new length. Our method is content-aware, requiring no tuning for different datasets, and fast, incurring negligible overhead. RLT yields a large speedup in training, reducing the wall-clock time to fine-tune a video transformer by 30% while matching baseline model performance. RLT also works without any training, increasing model throughput by 35% with only 0.1% drop in accuracy. RLT speeds up training at 30 FPS by more than 100%, and on longer video datasets, can reduce the token count by up to 80%. Our project page is at https://rccchoudhury.github.io/projects/rlt/.

Rohan Choudhury, JungEun Kim, Jinhyung Park et al.

Vision Transformers (ViTs) partition input images into uniformly sized patches regardless of their content, resulting in long input sequence lengths for high-resolution images. We present Adaptive Patch Transformers (APT), which addresses this by using multiple different patch sizes within the same image. APT reduces the total number of input tokens by allocating larger patch sizes in more homogeneous areas and smaller patches in more complex ones. APT achieves a drastic speedup in ViT inference and training, increasing throughput by 40% on ViT-L and 50% on ViT-H while maintaining downstream performance, and can be applied to a previously fine-tuned ViT, converging in as little as 1 epoch. It also significantly reduces training and inference time without loss of performance in high-resolution dense visual tasks, achieving up to 30\% faster training and inference in visual QA, object detection, and semantic segmentation.

Xingyu Liu, Kris M. Kitani

Manipulating articulated objects requires multiple robot arms in general. It is challenging to enable multiple robot arms to collaboratively complete manipulation tasks on articulated objects. In this paper, we present $\textbf{V-MAO}$, a framework for learning multi-arm manipulation of articulated objects. Our framework includes a variational generative model that learns contact point distribution over object rigid parts for each robot arm. The training signal is obtained from interaction with the simulation environment which is enabled by planning and a novel formulation of object-centric control for articulated objects. We deploy our framework in a customized MuJoCo simulation environment and demonstrate that our framework achieves a high success rate on six different objects and two different robots. We also show that generative modeling can effectively learn the contact point distribution on articulated objects.

Qichen Fu, Xingyu Liu, Kris M. Kitani

A key component of understanding hand-object interactions is the ability to identify the active object -- the object that is being manipulated by the human hand. In order to accurately localize the active object, any method must reason using information encoded by each image pixel, such as whether it belongs to the hand, the object, or the background. To leverage each pixel as evidence to determine the bounding box of the active object, we propose a pixel-wise voting function. Our pixel-wise voting function takes an initial bounding box as input and produces an improved bounding box of the active object as output. The voting function is designed so that each pixel inside of the input bounding box votes for an improved bounding box, and the box with the majority vote is selected as the output. We call the collection of bounding boxes generated inside of the voting function, the Relational Box Field, as it characterizes a field of bounding boxes defined in relationship to the current bounding box. While our voting function is able to improve the bounding box of the active object, one round of voting is typically not enough to accurately localize the active object. Therefore, we repeatedly apply the voting function to sequentially improve the location of the bounding box. However, since it is known that repeatedly applying a one-step predictor (i.e., auto-regressive processing with our voting function) can cause a data distribution shift, we mitigate this issue using reinforcement learning (RL). We adopt standard RL to learn the voting function parameters and show that it provides a meaningful improvement over a standard supervised learning approach. We perform experiments on two large-scale datasets: 100DOH and MECCANO, improving AP50 performance by 8% and 30%, respectively, over the state of the art.

Xingyu Liu, Shun Iwase, Kris M. Kitani

We present KDFNet, a novel method for 6D object pose estimation from RGB images. To handle occlusion, many recent works have proposed to localize 2D keypoints through pixel-wise voting and solve a Perspective-n-Point (PnP) problem for pose estimation, which achieves leading performance. However, such voting process is direction-based and cannot handle long and thin objects where the direction intersections cannot be robustly found. To address this problem, we propose a novel continuous representation called Keypoint Distance Field (KDF) for projected 2D keypoint locations. Formulated as a 2D array, each element of the KDF stores the 2D Euclidean distance between the corresponding image pixel and a specified projected 2D keypoint. We use a fully convolutional neural network to regress the KDF for each keypoint. Using this KDF encoding of projected object keypoint locations, we propose to use a distance-based voting scheme to localize the keypoints by calculating circle intersections in a RANSAC fashion. We validate the design choices of our framework by extensive ablation experiments. Our proposed method achieves state-of-the-art performance on Occlusion LINEMOD dataset with an average ADD(-S) accuracy of 50.3% and TOD dataset mug subset with an average ADD accuracy of 75.72%. Extensive experiments and visualizations demonstrate that the proposed method is able to robustly estimate the 6D pose in challenging scenarios including occlusion.

Xingyu Liu, Shun Iwase, Kris M. Kitani

We present a large-scale stereo RGB image object pose estimation dataset named the $\textbf{StereOBJ-1M}$ dataset. The dataset is designed to address challenging cases such as object transparency, translucency, and specular reflection, in addition to the common challenges of occlusion, symmetry, and variations in illumination and environments. In order to collect data of sufficient scale for modern deep learning models, we propose a novel method for efficiently annotating pose data in a multi-view fashion that allows data capturing in complex and flexible environments. Fully annotated with 6D object poses, our dataset contains over 393K frames and over 1.5M annotations of 18 objects recorded in 182 scenes constructed in 11 different environments. The 18 objects include 8 symmetric objects, 7 transparent objects, and 8 reflective objects. We benchmark two state-of-the-art pose estimation frameworks on StereOBJ-1M as baselines for future work. We also propose a novel object-level pose optimization method for computing 6D pose from keypoint predictions in multiple images. Project website: https://sites.google.com/view/stereobj-1m.

S. Alireza Golestaneh, Saba Dadsetan, Kris M. Kitani

The goal of No-Reference Image Quality Assessment (NR-IQA) is to estimate the perceptual image quality in accordance with subjective evaluations, it is a complex and unsolved problem due to the absence of the pristine reference image. In this paper, we propose a novel model to address the NR-IQA task by leveraging a hybrid approach that benefits from Convolutional Neural Networks (CNNs) and self-attention mechanism in Transformers to extract both local and non-local features from the input image. We capture local structure information of the image via CNNs, then to circumvent the locality bias among the extracted CNNs features and obtain a non-local representation of the image, we utilize Transformers on the extracted features where we model them as a sequential input to the Transformer model. Furthermore, to improve the monotonicity correlation between the subjective and objective scores, we utilize the relative distance information among the images within each batch and enforce the relative ranking among them. Last but not least, we observe that the performance of NR-IQA models degrades when we apply equivariant transformations (e.g. horizontal flipping) to the inputs. Therefore, we propose a method that leverages self-consistency as a source of self-supervision to improve the robustness of NRIQA models. Specifically, we enforce self-consistency between the outputs of our quality assessment model for each image and its transformation (horizontally flipped) to utilize the rich self-supervisory information and reduce the uncertainty of the model. To demonstrate the effectiveness of our work, we evaluate it on seven standard IQA datasets (both synthetic and authentic) and show that our model achieves state-of-the-art results on various datasets.

Xiaofang Wang, Shengcao Cao, Mengtian Li et al.

Existing neural architecture search (NAS) methods often return an architecture with good search performance but generalizes poorly to the test setting. To achieve better generalization, we propose a novel neighborhood-aware NAS formulation to identify flat-minima architectures in the search space, with the assumption that flat minima generalize better than sharp minima. The phrase ``flat-minima architecture'' refers to architectures whose performance is stable under small perturbations in the architecture (e.g., replacing a convolution with a skip connection). Our formulation takes the ``flatness'' of an architecture into account by aggregating the performance over the neighborhood of this architecture. We demonstrate a principled way to apply our formulation to existing search algorithms, including sampling-based algorithms and gradient-based algorithms. To facilitate the application to gradient-based algorithms, we also propose a differentiable representation for the neighborhood of architectures. Based on our formulation, we propose neighborhood-aware random search (NA-RS) and neighborhood-aware differentiable architecture search (NA-DARTS). Notably, by simply augmenting DARTS with our formulation, NA-DARTS outperforms DARTS and achieves state-of-the-art performance on established benchmarks, including CIFAR-10, CIFAR-100 and ImageNet.

Xiaofang Wang, Dan Kondratyuk, Eric Christiansen et al.

Committee-based models (ensembles or cascades) construct models by combining existing pre-trained ones. While ensembles and cascades are well-known techniques that were proposed before deep learning, they are not considered a core building block of deep model architectures and are rarely compared to in recent literature on developing efficient models. In this work, we go back to basics and conduct a comprehensive analysis of the efficiency of committee-based models. We find that even the most simplistic method for building committees from existing, independently pre-trained models can match or exceed the accuracy of state-of-the-art models while being drastically more efficient. These simple committee-based models also outperform sophisticated neural architecture search methods (e.g., BigNAS). These findings hold true for several tasks, including image classification, video classification, and semantic segmentation, and various architecture families, such as ViT, EfficientNet, ResNet, MobileNetV2, and X3D. Our results show that an EfficientNet cascade can achieve a 5.4x speedup over B7 and a ViT cascade can achieve a 2.3x speedup over ViT-L-384 while being equally accurate.

Zhengyi Luo, Ryo Hachiuma, Ye Yuan et al.

We propose a method for incorporating object interaction and human body dynamics into the task of 3D ego-pose estimation using a head-mounted camera. We use a kinematics model of the human body to represent the entire range of human motion, and a dynamics model of the body to interact with objects inside a physics simulator. By bringing together object modeling, kinematics modeling, and dynamics modeling in a reinforcement learning (RL) framework, we enable object-aware 3D ego-pose estimation. We devise several representational innovations through the design of the state and action space to incorporate 3D scene context and improve pose estimation quality. We also construct a fine-tuning step to correct the drift and refine the estimated human-object interaction. This is the first work to estimate a physically valid 3D full-body interaction sequence with objects (e.g., chairs, boxes, obstacles) from egocentric videos. Experiments with both controlled and in-the-wild settings show that our method can successfully extract an object-conditioned 3D ego-pose sequence that is consistent with the laws of physics.

Zhengyi Luo, S. Alireza Golestaneh, Kris M. Kitani

We develop a technique for generating smooth and accurate 3D human pose and motion estimates from RGB video sequences. Our method, which we call Motion Estimation via Variational Autoencoder (MEVA), decomposes a temporal sequence of human motion into a smooth motion representation using auto-encoder-based motion compression and a residual representation learned through motion refinement. This two-step encoding of human motion captures human motion in two stages: a general human motion estimation step that captures the coarse overall motion, and a residual estimation that adds back person-specific motion details. Experiments show that our method produces both smooth and accurate 3D human pose and motion estimates.

S. Alireza Golestaneh, Kris M. Kitani

We address the task of active learning in the context of semantic segmentation and show that self-consistency can be a powerful source of self-supervision to greatly improve the performance of a data-driven model with access to only a small amount of labeled data. Self-consistency uses the simple observation that the results of semantic segmentation for a specific image should not change under transformations like horizontal flipping (i.e., the results should only be flipped). In other words, the output of a model should be consistent under equivariant transformations. The self-supervisory signal of self-consistency is particularly helpful during active learning since the model is prone to overfitting when there is only a small amount of labeled training data. In our proposed active learning framework, we iteratively extract small image patches that need to be labeled, by selecting image patches that have high uncertainty (high entropy) under equivariant transformations. We enforce pixel-wise self-consistency between the outputs of segmentation network for each image and its transformation (horizontally flipped) to utilize the rich self-supervisory information and reduce the uncertainty of the network. In this way, we are able to find the image patches over which the current model struggles the most to classify. By iteratively training over these difficult image patches, our experiments show that our active learning approach reaches $\sim96\%$ of the top performance of a model trained on all data, by using only $12\%$ of the total data on benchmark semantic segmentation datasets (e.g., CamVid and Cityscapes).

Xiaofang Wang, Xuehan Xiong, Maxim Neumann et al.

Convolutional operations have two limitations: (1) do not explicitly model where to focus as the same filter is applied to all the positions, and (2) are unsuitable for modeling long-range dependencies as they only operate on a small neighborhood. While both limitations can be alleviated by attention operations, many design choices remain to be determined to use attention, especially when applying attention to videos. Towards a principled way of applying attention to videos, we address the task of spatiotemporal attention cell search. We propose a novel search space for spatiotemporal attention cells, which allows the search algorithm to flexibly explore various design choices in the cell. The discovered attention cells can be seamlessly inserted into existing backbone networks, e.g., I3D or S3D, and improve video classification accuracy by more than 2% on both Kinetics-600 and MiT datasets. The discovered attention cells outperform non-local blocks on both datasets, and demonstrate strong generalization across different modalities, backbones, and datasets. Inserting our attention cells into I3D-R50 yields state-of-the-art performance on both datasets.

Yu-Jhe Li, Zhengyi Luo, Xinshuo Weng et al.

Person re-identification (re-ID) aims to recognize instances of the same person contained in multiple images taken across different cameras. Existing methods for re-ID tend to rely heavily on the assumption that both query and gallery images of the same person have the same clothing. Unfortunately, this assumption may not hold for datasets captured over long periods of time (e.g., weeks, months or years). To tackle the re-ID problem in the context of clothing changes, we propose a novel representation learning model which is able to generate a body shape feature representation without being affected by clothing color or patterns. We call our model the Color Agnostic Shape Extraction Network (CASE-Net). CASE-Net learns a representation of identity that depends only on body shape via adversarial learning and feature disentanglement. Due to the lack of large-scale re-ID datasets which contain clothing changes for the same person, we propose two synthetic datasets for evaluation. We create a rendered dataset SMPL-reID with different clothes patterns and a synthesized dataset Div-Market with different clothing color to simulate two types of clothing changes. The quantitative and qualitative results across 5 datasets (SMPL-reID, Div-Market, two benchmark re-ID datasets, a cross-modality re-ID dataset) confirm the robustness and superiority of our approach against several state-of-the-art approaches

Andrey Rudenko, Luigi Palmieri, Michael Herman et al.

With growing numbers of intelligent autonomous systems in human environments, the ability of such systems to perceive, understand and anticipate human behavior becomes increasingly important. Specifically, predicting future positions of dynamic agents and planning considering such predictions are key tasks for self-driving vehicles, service robots and advanced surveillance systems. This paper provides a survey of human motion trajectory prediction. We review, analyze and structure a large selection of work from different communities and propose a taxonomy that categorizes existing methods based on the motion modeling approach and level of contextual information used. We provide an overview of the existing datasets and performance metrics. We discuss limitations of the state of the art and outline directions for further research.

Jiaqi Guan, Ye Yuan, Kris M. Kitani et al.

Automatically reasoning about future human behaviors is a difficult problem but has significant practical applications to assistive systems. Part of this difficulty stems from learning systems' inability to represent all kinds of behaviors. Some behaviors, such as motion, are best described with continuous representations, whereas others, such as picking up a cup, are best described with discrete representations. Furthermore, human behavior is generally not fixed: people can change their habits and routines. This suggests these systems must be able to learn and adapt continuously. In this work, we develop an efficient deep generative model to jointly forecast a person's future discrete actions and continuous motions. On a large-scale egocentric dataset, EPIC-KITCHENS, we observe our method generates high-quality and diverse samples while exhibiting better generalization than related generative models. Finally, we propose a variant to continually learn our model from streaming data, observe its practical effectiveness, and theoretically justify its learning efficiency.

Aashi Manglik, Xinshuo Weng, Eshed Ohn-Bar et al.

We explore the possibility of using a single monocular camera to forecast the time to collision between a suitcase-shaped robot being pushed by its user and other nearby pedestrians. We develop a purely image-based deep learning approach that directly estimates the time to collision without the need of relying on explicit geometric depth estimates or velocity information to predict future collisions. While previous work has focused on detecting immediate collision in the context of navigating Unmanned Aerial Vehicles, the detection was limited to a binary variable (i.e., collision or no collision). We propose a more fine-grained approach to collision forecasting by predicting the exact time to collision in terms of milliseconds, which is more helpful for collision avoidance in the context of dynamic path planning. To evaluate our method, we have collected a novel dataset of over 13,000 indoor video segments each showing a trajectory of at least one person ending in a close proximity (a near collision) with the camera mounted on a mobile suitcase-shaped platform. Using this dataset, we do extensive experimentation on different temporal windows as input using an exhaustive list of state-of-the-art convolutional neural networks (CNNs). Our results show that our proposed multi-stream CNN is the best model for predicting time to near-collision. The average prediction error of our time to near collision is 0.75 seconds across the test videos.

Rawal Khirodkar, Kris M. Kitani

Domain Randomization (DR) is known to require a significant amount of training data for good performance. We argue that this is due to DR's strategy of random data generation using a uniform distribution over simulation parameters, as a result, DR often generates samples which are uninformative for the learner. In this work, we theoretically analyze DR using ideas from multi-source domain adaptation. Based on our findings, we propose Adversarial Domain Randomization (ADR) as an efficient variant of DR which generates adversarial samples with respect to the learner during training. We implement ADR as a policy whose action space is the quantized simulation parameter space. At each iteration, the policy's action generates labeled data and the reward is set as negative of learner's loss on this data. As a result, we observe ADR frequently generates novel samples for the learner like truncated and occluded objects for object detection and confusing classes for image classification. We perform evaluations on datasets like CLEVR, Syn2Real, and VIRAT for various tasks where we demonstrate that ADR outperforms DR by generating fewer data samples.

Rawal Khirodkar, Donghyun Yoo, Kris M. Kitani

We address the issue of domain gap when making use of synthetic data to train a scene-specific object detector and pose estimator. While previous works have shown that the constraints of learning a scene-specific model can be leveraged to create geometrically and photometrically consistent synthetic data, care must be taken to design synthetic content which is as close as possible to the real-world data distribution. In this work, we propose to solve domain gap through the use of appearance randomization to generate a wide range of synthetic objects to span the space of realistic images for training. An ablation study of our results is presented to delineate the individual contribution of different components in the randomization process. We evaluate our method on VIRAT, UA-DETRAC, EPFL-Car datasets, where we demonstrate that using scene specific domain randomized synthetic data is better than fine-tuning off-the-shelf models on limited real data.

Arjun Sharma, Mohit Sharma, Nicholas Rhinehart et al.

The use of imitation learning to learn a single policy for a complex task that has multiple modes or hierarchical structure can be challenging. In fact, previous work has shown that when the modes are known, learning separate policies for each mode or sub-task can greatly improve the performance of imitation learning. In this work, we discover the interaction between sub-tasks from their resulting state-action trajectory sequences using a directed graphical model. We propose a new algorithm based on the generative adversarial imitation learning framework which automatically learns sub-task policies from unsegmented demonstrations. Our approach maximizes the directed information flow in the graphical model between sub-task latent variables and their generated trajectories. We also show how our approach connects with the existing Options framework, which is commonly used to learn hierarchical policies.

Xiang Xu, Xiaofang Wang, Kris M. Kitani

We propose to use the concept of the Hamming bound to derive the optimal criteria for learning hash codes with a deep network. In particular, when the number of binary hash codes (typically the number of image categories) and code length are known, it is possible to derive an upper bound on the minimum Hamming distance between the hash codes. This upper bound can then be used to define the loss function for learning hash codes. By encouraging the margin (minimum Hamming distance) between the hash codes of different image categories to match the upper bound, we are able to learn theoretically optimal hash codes. Our experiments show that our method significantly outperforms competing deep learning-based approaches and obtains top performance on benchmark datasets.

Xinlei Pan, Eshed Ohn-Bar, Nicholas Rhinehart et al.

Humans are able to understand and perform complex tasks by strategically structuring the tasks into incremental steps or subgoals. For a robot attempting to learn to perform a sequential task with critical subgoal states, such states can provide a natural opportunity for interaction with a human expert. This paper analyzes the benefit of incorporating a notion of subgoals into Inverse Reinforcement Learning (IRL) with a Human-In-The-Loop (HITL) framework. The learning process is interactive, with a human expert first providing input in the form of full demonstrations along with some subgoal states. These subgoal states define a set of subtasks for the learning agent to complete in order to achieve the final goal. The learning agent queries for partial demonstrations corresponding to each subtask as needed when the agent struggles with the subtask. The proposed Human Interactive IRL (HI-IRL) framework is evaluated on several discrete path-planning tasks. We demonstrate that subgoal-based interactive structuring of the learning task results in significantly more efficient learning, requiring only a fraction of the demonstration data needed for learning the underlying reward function with the baseline IRL model.

Tanmay Shankar, Nicholas Rhinehart, Katharina Muelling et al.

We explore the problem of learning to decompose spatial tasks into segments, as exemplified by the problem of a painting robot covering a large object. Inspired by the ability of classical decision tree algorithms to construct structured partitions of their input spaces, we formulate the problem of decomposing objects into segments as a parsing approach. We make the insight that the derivation of a parse-tree that decomposes the object into segments closely resembles a decision tree constructed by ID3, which can be done when the ground-truth available. We learn to imitate an expert parsing oracle, such that our neural parser can generalize to parse natural images without ground truth. We introduce a novel deterministic policy gradient update, DRAG (i.e., DeteRministically AGgrevate) in the form of a deterministic actor-critic variant of AggreVaTeD, to train our neural parser. From another perspective, our approach is a variant of the Deterministic Policy Gradient suitable for the imitation learning setting. The deterministic policy representation offered by training our neural parser with DRAG allows it to outperform state of the art imitation and reinforcement learning approaches.

Sima Behpour, Kris M. Kitani, Brian D. Ziebart

The use of random perturbations of ground truth data, such as random translation or scaling of bounding boxes, is a common heuristic used for data augmentation that has been shown to prevent overfitting and improve generalization. Since the design of data augmentation is largely guided by reported best practices, it is difficult to understand if those design choices are optimal. To provide a more principled perspective, we develop a game-theoretic interpretation of data augmentation in the context of object detection. We aim to find an optimal adversarial perturbations of the ground truth data (i.e., the worst case perturbations) that forces the object bounding box predictor to learn from the hardest distribution of perturbed examples for better test-time performance. We establish that the game theoretic solution, the Nash equilibrium, provides both an optimal predictor and optimal data augmentation distribution. We show that our adversarial method of training a predictor can significantly improve test time performance for the task of object detection. On the ImageNet object detection task, our adversarial approach improves performance by over 16\% compared to the best performing data augmentation method

Donghyun Yoo, Haoqi Fan, Vishnu Naresh Boddeti et al.

Feature representations from pre-trained deep neural networks have been known to exhibit excellent generalization and utility across a variety of related tasks. Fine-tuning is by far the simplest and most widely used approach that seeks to exploit and adapt these feature representations to novel tasks with limited data. Despite the effectiveness of fine-tuning, itis often sub-optimal and requires very careful optimization to prevent severe over-fitting to small datasets. The problem of sub-optimality and over-fitting, is due in part to the large number of parameters used in a typical deep convolutional neural network. To address these problems, we propose a simple yet effective regularization method for fine-tuning pre-trained deep networks for the task of k-shot learning. To prevent overfitting, our key strategy is to cluster the model parameters while ensuring intra-cluster similarity and inter-cluster diversity of the parameters, effectively regularizing the dimensionality of the parameter search space. In particular, we identify groups of neurons within each layer of a deep network that shares similar activation patterns. When the network is to be fine-tuned for a classification task using only k examples, we propagate a single gradient to all of the neuron parameters that belong to the same group. The grouping of neurons is non-trivial as neuron activations depend on the distribution of the input data. To efficiently search for optimal groupings conditioned on the input data, we propose a reinforcement learning search strategy using recurrent networks to learn the optimal group assignments for each network layer. Experimental results show that our method can be easily applied to several popular convolutional neural networks and improve upon other state-of-the-art fine-tuning based k-shot learning strategies by more than10%

Arun Venkatraman, Nicholas Rhinehart, Wen Sun et al.

Recurrent neural networks (RNNs) are a vital modeling technique that rely on internal states learned indirectly by optimization of a supervised, unsupervised, or reinforcement training loss. RNNs are used to model dynamic processes that are characterized by underlying latent states whose form is often unknown, precluding its analytic representation inside an RNN. In the Predictive-State Representation (PSR) literature, latent state processes are modeled by an internal state representation that directly models the distribution of future observations, and most recent work in this area has relied on explicitly representing and targeting sufficient statistics of this probability distribution. We seek to combine the advantages of RNNs and PSRs by augmenting existing state-of-the-art recurrent neural networks with Predictive-State Decoders (PSDs), which add supervision to the network's internal state representation to target predicting future observations. Predictive-State Decoders are simple to implement and easily incorporated into existing training pipelines via additional loss regularization. We demonstrate the effectiveness of PSDs with experimental results in three different domains: probabilistic filtering, Imitation Learning, and Reinforcement Learning. In each, our method improves statistical performance of state-of-the-art recurrent baselines and does so with fewer iterations and less data.

Mohit Sharma, Kris M. Kitani, Joachim Groeger

We make an important connection to existing results in econometrics to describe an alternative formulation of inverse reinforcement learning (IRL). In particular, we describe an algorithm using Conditional Choice Probabilities (CCP), which are maximum likelihood estimates of the policy estimated from expert demonstrations, to solve the IRL problem. Using the language of structural econometrics, we re-frame the optimal decision problem and introduce an alternative representation of value functions due to (Hotz and Miller 1993). In addition to presenting the theoretical connections that bridge the IRL literature between Economics and Robotics, the use of CCPs also has the practical benefit of reducing the computational cost of solving the IRL problem. Specifically, under the CCP representation, we show how one can avoid repeated calls to the dynamic programming subroutine typically used in IRL. We show via extensive experimentation on standard IRL benchmarks that CCP-IRL is able to outperform MaxEnt-IRL, with as much as a 5x speedup and without compromising on the quality of the recovered reward function.

Anubhav Ashok, Nicholas Rhinehart, Fares Beainy et al.

While bigger and deeper neural network architectures continue to advance the state-of-the-art for many computer vision tasks, real-world adoption of these networks is impeded by hardware and speed constraints. Conventional model compression methods attempt to address this problem by modifying the architecture manually or using pre-defined heuristics. Since the space of all reduced architectures is very large, modifying the architecture of a deep neural network in this way is a difficult task. In this paper, we tackle this issue by introducing a principled method for learning reduced network architectures in a data-driven way using reinforcement learning. Our approach takes a larger `teacher' network as input and outputs a compressed `student' network derived from the `teacher' network. In the first stage of our method, a recurrent policy network aggressively removes layers from the large `teacher' model. In the second stage, another recurrent policy network carefully reduces the size of each remaining layer. The resulting network is then evaluated to obtain a reward -- a score based on the accuracy and compression of the network. Our approach uses this reward signal with policy gradients to train the policies to find a locally optimal student network. Our experiments show that we can achieve compression rates of more than 10x for models such as ResNet-34 while maintaining similar performance to the input `teacher' network. We also present a valuable transfer learning result which shows that policies which are pre-trained on smaller `teacher' networks can be used to rapidly speed up training on larger `teacher' networks.

Ryo Yonetani, Vishnu Naresh Boddeti, Kris M. Kitani et al.

We propose a privacy-preserving framework for learning visual classifiers by leveraging distributed private image data. This framework is designed to aggregate multiple classifiers updated locally using private data and to ensure that no private information about the data is exposed during and after its learning procedure. We utilize a homomorphic cryptosystem that can aggregate the local classifiers while they are encrypted and thus kept secret. To overcome the high computational cost of homomorphic encryption of high-dimensional classifiers, we (1) impose sparsity constraints on local classifier updates and (2) propose a novel efficient encryption scheme named doubly-permuted homomorphic encryption (DPHE) which is tailored to sparse high-dimensional data. DPHE (i) decomposes sparse data into its constituent non-zero values and their corresponding support indices, (ii) applies homomorphic encryption only to the non-zero values, and (iii) employs double permutations on the support indices to make them secret. Our experimental evaluation on several public datasets shows that the proposed approach achieves comparable performance against state-of-the-art visual recognition methods while preserving privacy and significantly outperforms other privacy-preserving methods.

Nicholas Rhinehart, Kris M. Kitani

We address the problem of incrementally modeling and forecasting long-term goals of a first-person camera wearer: what the user will do, where they will go, and what goal they seek. In contrast to prior work in trajectory forecasting, our algorithm, DARKO, goes further to reason about semantic states (will I pick up an object?), and future goal states that are far in terms of both space and time. DARKO learns and forecasts from first-person visual observations of the user's daily behaviors via an Online Inverse Reinforcement Learning (IRL) approach. Classical IRL discovers only the rewards in a batch setting, whereas DARKO discovers the states, transitions, rewards, and goals of a user from streaming data. Among other results, we show DARKO forecasts goals better than competing methods in both noisy and ideal settings, and our approach is theoretically and empirically no-regret.

Xiaofang Wang, Yi Shi, Kris M. Kitani

Hashing is one of the most popular and powerful approximate nearest neighbor search techniques for large-scale image retrieval. Most traditional hashing methods first represent images as off-the-shelf visual features and then produce hashing codes in a separate stage. However, off-the-shelf visual features may not be optimally compatible with the hash code learning procedure, which may result in sub-optimal hash codes. Recently, deep hashing methods have been proposed to simultaneously learn image features and hash codes using deep neural networks and have shown superior performance over traditional hashing methods. Most deep hashing methods are given supervised information in the form of pairwise labels or triplet labels. The current state-of-the-art deep hashing method DPSH~\cite{li2015feature}, which is based on pairwise labels, performs image feature learning and hash code learning simultaneously by maximizing the likelihood of pairwise similarities. Inspired by DPSH~\cite{li2015feature}, we propose a triplet label based deep hashing method which aims to maximize the likelihood of the given triplet labels. Experimental results show that our method outperforms all the baselines on CIFAR-10 and NUS-WIDE datasets, including the state-of-the-art method DPSH~\cite{li2015feature} and all the previous triplet label based deep hashing methods.

Yubo Zhang, Vishnu Naresh Boddeti, Kris M. Kitani

Accurately identifying hands in images is a key sub-task for human activity understanding with wearable first-person point-of-view cameras. Traditional hand segmentation approaches rely on a large corpus of manually labeled data to generate robust hand detectors. However, these approaches still face challenges as the appearance of the hand varies greatly across users, tasks, environments or illumination conditions. A key observation in the case of many wearable applications and interfaces is that, it is only necessary to accurately detect the user's hands in a specific situational context. Based on this observation, we introduce an interactive approach to learn a person-specific hand segmentation model that does not require any manually labeled training data. Our approach proceeds in two steps, an interactive bootstrapping step for identifying moving hand regions, followed by learning a personalized user specific hand appearance model. Concretely, our approach uses two convolutional neural networks: (1) a gesture network that uses pre-defined motion information to detect the hand region; and (2) an appearance network that learns a person specific model of the hand region based on the output of the gesture network. During training, to make the appearance network robust to errors in the gesture network, the loss function of the former network incorporates the confidence of the gesture network while learning. Experiments demonstrate the robustness of our approach with an F1 score over 0.8 on all challenging datasets across a wide range of illumination and hand appearance variations, improving over a baseline approach by over 10%.

Xiaofang Wang, Kris M. Kitani, Martial Hebert

Measuring visual similarity is critical for image understanding. But what makes two images similar? Most existing work on visual similarity assumes that images are similar because they contain the same object instance or category. However, the reason why images are similar is much more complex. For example, from the perspective of category, a black dog image is similar to a white dog image. However, in terms of color, a black dog image is more similar to a black horse image than the white dog image. This example serves to illustrate that visual similarity is ambiguous but can be made precise when given an explicit contextual perspective. Based on this observation, we propose the concept of contextual visual similarity. To be concrete, we examine the concept of contextual visual similarity in the application domain of image search. Instead of providing only a single image for image similarity search (\eg, Google image search), we require three images. Given a query image, a second positive image and a third negative image, dissimilar to the first two images, we define a contextualized similarity search criteria. In particular, we learn feature weights over all the feature dimensions of each image such that the distance between the query image and the positive image is small and their distances to the negative image are large after reweighting their features. The learned feature weights encode the contextualized visual similarity specified by the user and can be used for attribute specific image search. We also show the usefulness of our contextualized similarity weighting scheme for different tasks, such as answering visual analogy questions and unsupervised attribute discovery.

Jonathan Shen, Noranart Vesdapunt, Vishnu N. Boddeti et al.

Deep convolutional neural networks continue to advance the state-of-the-art in many domains as they grow bigger and more complex. It has been observed that many of the parameters of a large network are redundant, allowing for the possibility of learning a smaller network that mimics the outputs of the large network through a process called Knowledge Distillation. We show, however, that standard Knowledge Distillation is not effective for learning small models for the task of pedestrian detection. To improve this process, we introduce a higher-dimensional hint layer to increase information flow. We also estimate the variance in the outputs of the large network and propose a loss function to incorporate this uncertainty. Finally, we attempt to boost the complexity of the small network without increasing its size by using as input hand-designed features that have been demonstrated to be effective for pedestrian detection. We succeed in training a model that contains $400\times$ fewer parameters than the large network while outperforming AlexNet on the Caltech Pedestrian Dataset.

Ryo Yonetani, Kris M. Kitani, Yoichi Sato

We envision a future time when wearable cameras are worn by the masses and recording first-person point-of-view videos of everyday life. While these cameras can enable new assistive technologies and novel research challenges, they also raise serious privacy concerns. For example, first-person videos passively recorded by wearable cameras will necessarily include anyone who comes into the view of a camera -- with or without consent. Motivated by these benefits and risks, we developed a self-search technique tailored to first-person videos. The key observation of our work is that the egocentric head motion of a target person (ie, the self) is observed both in the point-of-view video of the target and observer. The motion correlation between the target person's video and the observer's video can then be used to identify instances of the self uniquely. We incorporate this feature into the proposed approach that computes the motion correlation over densely-sampled trajectories to search for a target individual in observer videos. Our approach significantly improves self-search performance over several well-known face detectors and recognizers. Furthermore, we show how our approach can enable several practical applications such as privacy filtering, target video retrieval, and social group clustering.

Minghuang Ma, Haoqi Fan, Kris M. Kitani

We bring together ideas from recent work on feature design for egocentric action recognition under one framework by exploring the use of deep convolutional neural networks (CNN). Recent work has shown that features such as hand appearance, object attributes, local hand motion and camera ego-motion are important for characterizing first-person actions. To integrate these ideas under one framework, we propose a twin stream network architecture, where one stream analyzes appearance information and the other stream analyzes motion information. Our appearance stream encodes prior knowledge of the egocentric paradigm by explicitly training the network to segment hands and localize objects. By visualizing certain neuron activation of our network, we show that our proposed architecture naturally learns features that capture object attributes and hand-object configurations. Our extensive experiments on benchmark egocentric action datasets show that our deep architecture enables recognition rates that significantly outperform state-of-the-art techniques -- an average $6.6\%$ increase in accuracy over all datasets. Furthermore, by learning to recognize objects, actions and activities jointly, the performance of individual recognition tasks also increase by $30\%$ (actions) and $14\%$ (objects). We also include the results of extensive ablative analysis to highlight the importance of network design decisions..

Nicholas Rhinehart, Kris M. Kitani

When people observe and interact with physical spaces, they are able to associate functionality to regions in the environment. Our goal is to automate dense functional understanding of large spaces by leveraging sparse activity demonstrations recorded from an ego-centric viewpoint. The method we describe enables functionality estimation in large scenes where people have behaved, as well as novel scenes where no behaviors are observed. Our method learns and predicts "Action Maps", which encode the ability for a user to perform activities at various locations. With the usage of an egocentric camera to observe human activities, our method scales with the size of the scene without the need for mounting multiple static surveillance cameras and is well-suited to the task of observing activities up-close. We demonstrate that by capturing appearance-based attributes of the environment and associating these attributes with activity demonstrations, our proposed mathematical framework allows for the prediction of Action Maps in new environments. Additionally, we offer a preliminary glance of the applicability of Action Maps by demonstrating a proof-of-concept application in which they are used in concert with activity detections to perform localization.

Wei-Chiu Ma, De-An Huang, Namhoon Lee et al.

We develop predictive models of pedestrian dynamics by encoding the coupled nature of multi-pedestrian interaction using game theory, and deep learning-based visual analysis to estimate person-specific behavior parameters. Building predictive models for multi-pedestrian interactions however, is very challenging due to two reasons: (1) the dynamics of interaction are complex interdependent processes, where the predicted behavior of one pedestrian can affect the actions taken by others and (2) dynamics are variable depending on an individuals physical characteristics (e.g., an older person may walk slowly while the younger person may walk faster). To address these challenges, we (1) utilize concepts from game theory to model the interdependent decision making process of multiple pedestrians and (2) use visual classifiers to learn a mapping from pedestrian appearance to behavior parameters. We evaluate our proposed model on several public multiple pedestrian interaction video datasets. Results show that our strategic planning model explains human interactions 25% better when compared to state-of-the-art methods.